For a function approximation problem I'm trying to accumulate gradients but I find that sometimes some of these gradients are nan(i.e. undefined) even though the loss is always real. I think this might be due to numerical instabilities and I'm basically looking for a simple method for removing the nans from the computed gradients.
Starting with the solution to this question I tried doing the following:
# Optimizer definition - nothing different from any classical example
opt = tf.train.AdamOptimizer()
## Retrieve all trainable variables you defined in your graph
tvs = tf.trainable_variables()
## Creation of a list of variables with the same shape as the trainable ones
# initialized with 0s
accum_vars = [tf.Variable(tf.zeros_like(tv.initialized_value()), trainable=False) for tv in tvs]
zero_ops = [tv.assign(tf.zeros_like(tv)) for tv in accum_vars]
## Calls the compute_gradients function of the optimizer to obtain... the list of gradients
gvs_ = opt.compute_gradients(rmse, tvs)
gvs =tf.where(tf.is_nan(gvs_), tf.zeros_like(gvs_), gvs_)
## Adds to each element from the list you initialized earlier with zeros its gradient (works because accum_vars and gvs are in the same order)
accum_ops = [accum_vars[i].assign_add(gv[0]) for i, gv in enumerate(gvs)]
## Define the training step (part with variable value update)
train_step = opt.apply_gradients([(accum_vars[i], gv[1]) for i, gv in enumerate(gvs)])
So basically, the key idea is this line:
gvs =tf.where(tf.is_nan(gvs_), tf.zeros_like(gvs_), gvs_)
But when I apply this idea I obtain the following error:
ValueError: Tried to convert 'x' to a tensor and failed. Error:
Dimension 1 in both shapes must be equal, but are 30 and 9. Shapes are
[2,30] and [2,9]. From merging shape 2 with other shapes. for
'IsNan/packed' (op: 'Pack') with input shapes: [2,9,30], [2,30,9],
[2,30], [2,9].
compute_gradients returns a list of tensors in your case. You may want to do:
gvs_ = [(tf.where(tf.is_nan(grad), tf.zeros_like(grad), grad), val) for grad,val in gvs_]
Related
I have a pytorch model:
model = torch.nn.Sequential(
torch.nn.LSTM(40, 256, 3, batch_first=True),
torch.nn.Linear(256, 256),
torch.nn.ReLU()
)
And for the LSTM layer, I want to retrieve only the last hidden state from the batch to pass through the rest of the layers. Ex:
_, (hidden, _) = lstm(data)
hidden = hidden[-1]
Though, that example only works for a subclassed model. I need to somehow do this on a nn.Sequential() model that way when I save it, it can properly be converted to a tensorflow.js model. The reason I can't make and train this model in tensorflow.js is because I'm trying to implement this repo: Resemblyzer in tensorflow.js while still using the same weights as the pretrained Resemblyzer model which was made in pytorch as a subclassed model. I thought of using the torchvisions.transformations.Lambda() transformation but I would assume that would make it incompatible with tensorflow.js. Is there any way to make this possible while still allowing the model to convert properly?
You could split up your sequential but only doing so in the forward definition of your model on inference. Once defined:
model = nn.Sequential(nn.LSTM(40, 256, 3, batch_first=True),
nn.Linear(256, 256),
nn.ReLU())
You can split it:
>>> lstm, fc = model[0], model[1:]
Then infer in two steps:
>>> out, (hidden, _) = lstm(data)
>>> hidden = hidden[-1]
>>> out = fc(out) # <- or fc(out[-1]) depending on what you want
Though the answer is provided above, I thought of elaborating on the same as PyTorch LSTM documentation is confusing.
In TF, we directly get the last_state as the output. No further action needed.
Let us check the Torch output of LSTM:
There are 2 outputs - a sequence and a tuple. We are interested in the last state so we can ignore the sequence and focus on the tuple. The tuple consists of 2 values - the first is the hidden state of the last cell (of all layers in the LSTM) and the second is the cell state of the last cell (again of all layers in the LSTM). We are interested in the hidden state. So
_, tup = self.bilstm(inp)
We are interested in tup[0]. Let us dig further into this.
The shape of tup[0] is somewhat odd with batch size at the centre. On the left of the batch size is the number of layers in the LSTM (multiply 2 if is biLSTM). On the right is the dimension you have provided while defining the LSTM. You could take the output from the last layer by simply doing a tup[0][-1] which is the answer provided above.
Alternatively if you want to make use of hidden states across layers, you may try something like:
out = tup[0].swapaxes(0,1)
out = out.reshape(*out.shape[:-2], -1)
The first line produces shape of batch_size, num_layers, hidden_size_specified. The second line produces shape of batch_size, num_layers x hidden_size_specified
(For e.g., Let us say, yours is a biLSTM and you have 3 layers and your hiddensize is 100, you could choose to concatenate the output such that you get one vector of 2 x 3 x 100 = 600 dimensions and then run a simple linear layer on top of this to get the output you want.)
There is another way to get the output of the LSTM. We discussed that the first output of an LSTM is a sequence:
sequence, tup = self.bilstm(inp)
This sequence is the output of the LAST hidden layer of the LSTM. It is a sequence because it contains hidden states of EVERY cell in this layer. So its length will be the input sequence length that you have provided. We could choose to take the hidden state of the last element in the sequence by doing a:
#shape of sequence is: batch_size, seq_size, dim
sequence = sequence.swapaxes(0,1)
#shape of sequence is: seq_size, batch_size, dim
sequence = sequence[-1]
#shape of sequence is: batch_size, dim (ie last seq is taken)
Needless to say this will be the same value we got by taking the last layer from tup[0]. Well, not quite! If the LSTM is a biLSTM, then using the sequence approach returns is 2 x hidden_size dim output (which is correct) wheras using the tup[0][-1] approach will give us only hidden_size dim even for a biLSTM. OP's LSTM is a non-biLSTM so both answers hold true.
I'm trying to create a matrix of gradient with the gradient of each observation by parameters and EPOCH. If my model has 100 obs, 1000 params and 10 Epoch, my matrix should be (100,1000,10).
The problem is that I'm not able to get those gradient. The parameters and the observation are set at required_gradient=True.
I've tried to run this after each observation pass thru the net:
for p in net.parameters():
paramgradlist.append(p.grad)
But the gradient stays the same of each params stays the same for all observations.
Thank you
You are not copying your data and instead of storing a reference to the gradients. In the end, this means all your observations will be the same (i.e. the gradients' final value).
Instead, you could clone the gradients before appending them to the list:
for p in net.parameters():
paramgradlist.append(p.grad.clone())
I'm working with time-variant graph embedding, where at each time step, the adjacency matrix of the graph changes. The main idea is to perform the node embedding of each timestep of the graph by looking to a set of node features and the adjacency matrix. The node embedding step is long and complicated, and is not part of the core of the problem, so I will skip this part. Suffice it to say that I use Graph Convolutional Network to embed the nodes.
Consider that I have a stack of B adjacency matrices A with sizes NxN, where B = batch size and N = number of nodes in the graph. Also, the matrices are stacked according to a time series, where matrix in index i comes before matrix in index i+1. I have already embedded the nodes of the graph, which results in a matrix of dimensions B x N x E, where E = size of the embedding (parameter). Note that the model has to deal with any graph, therefore, N is not a parameter. Another important comment is that each batch contains adjacency matrices from the same graph, and therefore all matrices of a batch have the same number of node, but the matrices of other batches may have different number of nodes.
I now need to pass these embedding through an LSTM cell. I never used Keras before, so I'm having a hard time making the Keras LSTM blend in my Tensorflow code. What I want to do is: pass each node embedding through an LSTM such that the number of timesteps = B and the LSTM batch size = N, that is, the input to my LSTM has the shape [N, B, E], where N and B are only known through execution time. I want the output of my LSTM to have the shape of [B, E*E]. The embedding matrix is called here self.embed_mat. Here is my code:
def _LSTM_layer(self):
with tf.variable_scope(self.scope, reuse=tf.AUTO_REUSE), tf.device(self.device):
in_shape = tf.shape(self.embed_mat)
lstm_input = tf.reshape(self.embed_mat, [in_shape[1], in_shape[0], EMBED_SIZE]) #lstm = [N, B, E]
input_plh = K.placeholder(name="lstm_input", shape=(None, None, EMBED_SIZE))
lstm = LSTM(EMBED_SIZE*EMBED_SIZE, input_shape=(None, None, EMBED_SIZE))
get_output = K.function(inputs=[input_plh], outputs=[lstm(input_plh)])
h = get_output([lstm_input])
I am a bit lost with the K.function part. All I want is the output tensor of the LSTM cell. I've seen that in order to get that with Keras, we need to use K.function, but I don't quite get it what it does. When I call get_output([lstm_input]), I get the following error:
tensorflow.python.framework.errors_impl.InvalidArgumentError: You must feed a value for placeholder tensor 'worker_global/A/shape' with dtype int64 and shape [?]
Here, A is the stacked adjacency matrices with dimension BxNxN. What is going on here? Does the value of N needs to be known during graph building step? I think I made some dumb mistake with the LSTM cell, but I can't get what it is.
Thanks in advance!
If you want to get the output of your LSTM layer "out" given input of "inp" in a keras Sequential() model called "model," where "inp" is your first / input layer and "out" is an LSTM layer that happens to be, for the sake of this example, in the 4th position in your sequential model, you would obtain the output of that LSTM layer from the data you call "lstm_input" above with the following code:
inp = model.layers[0].input
out = model.layers[3].output
inp_to_out = K.function([inp], [out])
output = inp_to_out([lstm_input])
I'm trying to randomly subsample the prediction and target array for my loss calculation.
idx = torch.randperm(target.shape[0])
target = target.index_select(0, idx[0, sample_size]
However I'm getting this error message.
index_select(): argument 'index' (position 2) must be Variable, not torch.LongTensor
Does anyone know how to fix this?
Edit:
I got one step closer. It seems like torch.randperm does not return a torch variable, so one has to explicitly convert the output:
idx = torch.randperm(target.shape[0])
idx = Variable(idx).cuda()
target = target.index_select(0, idx[0, sample_size]
only problem is now that the backpropagation fails. Seems like the operation of randomly subsampling is causing an issue with the dimensions.
However the dimensions seem to be fine when calculating the loss:
loss = F.nll_loss(prediction, target.view(-1)) # prediction shape is [Nx12] and target shape is N
Unfortunately when calling loss.backward() I get this error message:
RuntimeError: The expanded size of the tensor (12) must match the existing size (217456) at non-singleton dimension 1
This release of PyTorch seems provide the PackedSequence for variable lengths of input for recurrent neural network. However, I found it's a bit hard to use it correctly.
Using pad_packed_sequence to recover an output of a RNN layer which were fed by pack_padded_sequence, we got a T x B x N tensor outputs where T is the max time steps, B is the batch size and N is the hidden size. I found that for short sequences in the batch, the subsequent output will be all zeros.
Here are my questions.
For a single output task where the one would need the last output of all the sequences, simple outputs[-1] will give a wrong result since this tensor contains lots of zeros for short sequences. One will need to construct indices by sequence lengths to fetch the individual last output for all the sequences. Is there more simple way to do that?
For a multiple output task (e.g. seq2seq), usually one will add a linear layer N x O and reshape the batch outputs T x B x O into TB x O and compute the cross entropy loss with the true targets TB (usually integers in language model). In this situation, do these zeros in batch output matters?
Question 1 - Last Timestep
This is the code that i use to get the output of the last timestep. I don't know if there is a simpler solution. If it is, i'd like to know it. I followed this discussion and grabbed the relative code snippet for my last_timestep method. This is my forward.
class BaselineRNN(nn.Module):
def __init__(self, **kwargs):
...
def last_timestep(self, unpacked, lengths):
# Index of the last output for each sequence.
idx = (lengths - 1).view(-1, 1).expand(unpacked.size(0),
unpacked.size(2)).unsqueeze(1)
return unpacked.gather(1, idx).squeeze()
def forward(self, x, lengths):
embs = self.embedding(x)
# pack the batch
packed = pack_padded_sequence(embs, list(lengths.data),
batch_first=True)
out_packed, (h, c) = self.rnn(packed)
out_unpacked, _ = pad_packed_sequence(out_packed, batch_first=True)
# get the outputs from the last *non-masked* timestep for each sentence
last_outputs = self.last_timestep(out_unpacked, lengths)
# project to the classes using a linear layer
logits = self.linear(last_outputs)
return logits
Question 2 - Masked Cross Entropy Loss
Yes, by default the zero padded timesteps (targets) matter. However, it is very easy to mask them. You have two options, depending on the version of PyTorch that you use.
PyTorch 0.2.0: Now pytorch supports masking directly in the CrossEntropyLoss, with the ignore_index argument. For example, in language modeling or seq2seq, where i add zero padding, i mask the zero padded words (target) simply like this:
loss_function = nn.CrossEntropyLoss(ignore_index=0)
PyTorch 0.1.12 and older: In the older versions of PyTorch, masking was not supported, so you had to implement your own workaround. I solution that i used, was masked_cross_entropy.py, by jihunchoi. You may be also interested in this discussion.
A few days ago, I found this method which uses indexing to accomplish the same task with a one-liner.
I have my dataset batch first ([batch size, sequence length, features]), so for me:
unpacked_out = unpacked_out[np.arange(unpacked_out.shape[0]), lengths - 1, :]
where unpacked_out is the output of torch.nn.utils.rnn.pad_packed_sequence.
I have compared it with the method described here, which looks similar to the last_timestep() method Christos Baziotis is using above (also recommended here), and the results are the same in my case.